A circuit breaker fault diagnosis method and system based on a convolutional prototype network

Abstract

A circuit breaker fault diagnosis method and system based on a convolution prototype network, the method comprising: collecting circuit breaker fault data, and preprocessing the circuit breaker fault data to obtain a circuit breaker fault data training sample set; constructing a circuit breaker fault diagnosis model of the convolution prototype network, and inputting the training sample set to train the model; taking each type of prototype sample point as a clustering center, and taking a sample feature space distance as an optimization target of a convolution feature self-extraction network, so as to effectively improve the intra-class aggregation and inter-class dispersion of fault sample features, and improve the classification accuracy of the model on sample types; and inputting the circuit breaker fault data sample set into the trained circuit breaker fault diagnosis model of the convolution prototype network to obtain a fault diagnosis result.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of substation fault diagnosis, and particularly relates to a circuit breaker fault diagnosis method and system based on a convolution prototype network. BACKGROUND

[0002] High-voltage circuit breakers play a very key role in power grid stability control and fault isolation. The operation and maintenance of high-voltage circuit breakers is an important part of the daily work of substation maintenance personnel. With the continuous expansion of the power grid, the types and categories of circuit breakers are numerous, and the operating environment is complex. The traditional periodic maintenance has been unable to meet the requirements of the power grid for electrical equipment condition-based maintenance. Therefore, it is necessary to improve the monitoring level of the operating state of the circuit breaker. Based on theoretical analysis, relying on actual monitoring data, it is necessary to study more practical and efficient circuit breaker fault diagnosis technology to assist maintenance personnel to more accurately formulate maintenance strategies and improve maintenance efficiency, and to improve the security of the power grid.

[0003] At present, the research on high-voltage circuit breaker fault diagnosis mostly takes the opening and closing coil current waveform as the analysis object, and analyzes the current waveform characteristics in different stages of the opening and closing process to realize the judgment of the corresponding fault cause of the circuit breaker. The existing method well studies the analysis and diagnosis of known fault types of the circuit breaker. However, in fact, the number of fault samples in the circuit breaker historical data is small, and only through simulation or simulation experiments cannot master all the fault conditions of the circuit breaker, and the traditional learning mode based on a closed set only allocates a corresponding probability space to the fault types that have appeared in the training set, so that all fault samples will be attributed to the known category (appeared in the training set), and do not have the ability to distinguish and distinguish unknown (not appeared in the training set) fault conditions.

[0004] The existing file 1 (CN 106291351 B) is based on a high-voltage circuit breaker fault detection method of a convolutional neural network algorithm. A magnetic balance type Hall current sensor is connected with a circuit breaker closing and opening coil and a data processing system to construct a closing and opening coil current online monitoring system. The closing and opening coil current data obtained by real-time monitoring is used. A deep learning algorithm based on a convolutional neural network is used to construct a fault type prediction model. Part of the closing and opening coil current data is input into the constructed fault type prediction model for training. Part of the closing and opening coil current data is input into the trained fault type prediction model, and the input closing and opening coil current data is processed by the fault type prediction model to complete the high-voltage circuit breaker fault detection. The deficiency of the prior art file 1 is that it can only identify known fault types. For example, in the embodiment of the application, the fault output type in Table 2 is A, B, C, D and E. If the input sample is not one of the five types, the method in file 1 can still only output results within the five types. That is, the probability space has been divided into five types, and there is no corresponding probability space for unknown types. Therefore, the method of the existing file 1 cannot judge unknown fault types other than the five types. SUMMARY

[0005] To solve the problems in the prior art, the application provides a circuit breaker fault diagnosis method and system based on a convolutional prototype network. A classification function is constructed based on the clustering idea, prototype sample points of various types are used as clustering centers, and sample feature space distance is used as an optimization target of a convolutional feature self-extraction network to effectively improve the intra-class aggregation and inter-class dispersion of fault sample features. Meanwhile, probability space is allocated for unknown fault samples to realize the discrimination and differentiation ability for unknown circuit breaker faults.

[0006] The application adopts the following technical solutions.

[0007] A circuit breaker fault diagnosis method based on a convolutional prototype network comprises the following steps:

[0008] Step 1: Collecting circuit breaker fault data and pre-processing the circuit breaker fault data to obtain a circuit breaker fault data training sample set;

[0009] Step 2: Constructing a circuit breaker fault diagnosis model of a convolutional prototype network, inputting the training sample set to train the model, and using sample feature space distance as an optimization target of the circuit breaker fault diagnosis model of the convolutional prototype network;

[0010] Step 3: Inputting the circuit breaker fault data sample set into the trained circuit breaker fault diagnosis model of the convolutional prototype network to obtain a fault diagnosis result.

[0011] Preferably, the circuit breaker failure data in step 1 includes: coil aging data A, coil aging data B, coil aging data C, core sticking data A, and core sticking data B.

[0012] Preferably, the circuit breaker failure data preprocessing method in step 1 includes: mean method, EMD method, and wavelet decomposition method.

[0013] Preferably, the circuit breaker failure diagnosis model of the convolutional prototype network in step 2 includes: an input layer, a convolutional pooling layer 1, a convolutional pooling layer 2, and a full connection layer, a classification function is used to classify the output features of the convolutional layer, a class neighborhood is introduced to train the loss function, and the loss function is used to drive the update of the convolutional layer hyperparameters, and the model training process specifically includes:

[0014] Step 2.1, a classification function is constructed in a clustering method, and the training sample set of the circuit breaker failure data is classified, and if the matching degree of the sample with respect to a certain class is the highest, that is, the sample feature space distance is the smallest, then the class of the sample can be determined;

[0015] Step 2.2, the size of the matching degree is quantified in a probability form, and the classification function is improved;

[0016] Step 2.3, the sample feature space distance is estimated according to the improved classification function;

[0017] Step 2.4, a class neighborhood is introduced to train the loss function, and the feature clustering target directly drives the optimization of the circuit breaker failure diagnosis model of the convolutional prototype network for sample feature extraction.

[0018] Preferably, step 2.1 includes: the classification function is as shown in formulas (3) and (4):

[0019]

[0020]

[0021] In the formula,

[0022] C represents the total number of sample types,

[0023] i represents the sample classification class,

[0024] j represents the prototype sample point number of each class, j∈{1,2,…,K},

[0025] K represents the number of prototype samples selected for each type,

[0026] f(x) represents a convolutional prototype network function,

[0027] m ij represents a prototype sample point,

[0028] g i (x) represents the matching score of sample x and known class i;

[0029] For sample x, if it does not match all known classes, it is expressed as:

[0030]

[0031] In the formula,

[0032] T represents the matching degree threshold of sample features and prototype samples of each class.

[0033] Preferably, step 2.2 includes: the matching degree calculation method is converted to formula (6) shown as:

[0034]

[0035] In the formula,

[0036] represents the distance between two vectors,

[0037] p i (x) represents the matching score of sample x and known class i after probability quantization,

[0038] γ represents a hyperparameter for controlling the conversion of distance results to probability form, and takes a value of 0.001,

[0039] Rewrite formula (5) as:

[0040]

[0041] In the formula, T1 represents the matching degree threshold in the probability form.

[0042] Preferably, step 2.3 specifically includes: assuming that for sample x, i1 and i2 represent two samples closest to and second closest to sample x in the feature distance of a certain type, and j is a sample of a different class, then the feature distance relationship of the three samples is expressed as:

[0043] g i1 (x) > g i2 (x) >> g j (x) (8)

[0044] Then the approximate estimation formula of the sample feature distance is shown as follows:

[0045]

[0046] Therefore, the selection of the matching degree threshold T1 is obtained from formula (9).

[0047] Preferably, step 2.4 specifically comprises: in the definition of the training loss function, the training target of the multi-classification model is to make the sample belong to its correct class while not belonging to all other classes; in order to convert the distance into binary classification, i.e. whether it belongs to this class, the model sets a radius for each prototype sample point to form a class neighborhood, and the expression is as follows:

[0048]

[0049] In the formula,

[0050] f represents the result of the sample x after being extracted by the convolution network,

[0051] m ij represents the prototype sample point,

[0052] r represents the radius,

[0053] R i represents the decision region;

[0054] For each class i, the class neighborhood constitutes the decision region R i of the feature space, and the center point of the decision region, i.e. the prototype sample point, is the mean point of the features of the samples in the class and is constantly updated as the features of the samples change in the training;

[0055] For the sample x, f(x) ∈ R i indicates that x belongs to class i, otherwise it does not belong to class i, if there is a sample that does not satisfy f(x) ∈ R i for all types, it can be judged as an unknown class, and through a one-versus-all training strategy, the target is sample x, and the feature classification can be expressed as:

[0056]

[0057] In the formula,

[0058] R y represents the decision region of the sample class y,

[0059] R c represents the decision region of the sample class c,

[0060] The calculation process of the above binary classification is expressed as a function as follows:

[0061]

[0062]

[0063] In the formula,

[0064] h i (x) represents the function expression form of the same class sample i,

[0065] h y (x) represents the result of the sample class y,

[0066] h c (x) represents the result of the sample class c (c≠y) ;

[0067] In order to make the loss function more smooth, the binary discriminant function is converted into a probability function (denoted as σ) by using the sigmoid function, as follows:

[0068] σ i (x) = sigmoid (h i (x)) (14)

[0069] Then the loss function of the model is:

[0070]

[0071] In the formula,

[0072] σ i (x) represents the probability result of the same sample i,

[0073] σ y (x) represents the probability result of the sample class y,

[0074] σ c (x) represents the probability result of the sample class c (c≠y).

[0075] A circuit breaker fault diagnosis system based on a convolutional prototype network, comprising: a collection module, a modeling and training module, and a diagnosis module, wherein:

[0076] The collection module is used to collect circuit breaker fault data, and pre-process the circuit breaker fault data to obtain a circuit breaker fault data training sample set;

[0077] The modeling and training module is used to construct a circuit breaker fault diagnosis model of the convolutional prototype network, input the training sample set to train the model, and take the sample feature space distance as an optimization target of the circuit breaker fault diagnosis model of the convolutional prototype network;

[0078] The diagnosis module is used to input the circuit breaker fault data sample set into the trained circuit breaker fault diagnosis model of the convolutional prototype network to obtain a fault diagnosis result.

[0079] A terminal comprising a processor and a storage medium; wherein:

[0080] The storage medium is used to store instructions;

[0081] The processor is configured to operate according to the instructions to perform steps of a circuit breaker fault diagnosis method based on a convolution prototype network.

[0082] A computer readable storage medium having stored thereon a computer program, wherein the program, when executed by a processor, implements steps of a circuit breaker fault diagnosis method based on a convolution prototype network.

[0083] The present application has the advantages that, compared with the prior art,

[0084] The present application provides a circuit breaker fault diagnosis method based on a convolution prototype network. First, the fault data is preprocessed, and the normal data and the fault sample data are combined to highlight the difference of the fault, and then the traditional closed set learning model is improved, and the classification function is constructed based on the clustering idea to adapt to the processing of the open data set, so that the unknown circuit breaker fault type is diagnosed. At the same time, taking each type of prototype sample point as the clustering center, and taking the sample feature space distance as the optimization target of the convolution feature self-extraction network, the intra-class aggregation and inter-class dispersion of the fault sample features are effectively improved, and the classification accuracy of the model to the sample type is improved. In this paper, the simulation experimental data is taken as the analysis object, the accuracy of the proposed method in feature extraction and unknown type diagnosis is analyzed in detail, and the proposed method is compared with the existing method, and the superiority of the model in the feature extraction and classification stages is analyzed. BRIEF DESCRIPTION OF DRAWINGS

[0085] Figure 1 is a flow chart of the circuit breaker fault diagnosis method based on the convolution prototype network of the present application;

[0086] Figure 2 is a schematic diagram of a convolution prototype network;

[0087] Figure 3 is a closing current waveform diagram of a circuit breaker with a closing coil aging fault;

[0088] Figure 4 is a closing current waveform diagram of a circuit breaker with a core sticking fault;

[0089] Figure 5 is a waveform diagram of six fault samples of a high-voltage circuit breaker;

[0090] Figure 6 is a training process diagram of the model of the present application;

[0091] Figure 7 is a feature extraction result diagram of the model of the present application for training data;

[0092] Figure 8 is a feature extraction result diagram of a traditional CNN model for training data;

[0093] Figure 9 is a feature extraction effect evaluation comparison chart;

[0094] Figure 10 is a feature extraction result schematic diagram containing 5 types of training data;

[0095] Figure 11 is a feature extraction result schematic diagram of 6 types of test data. DETAILED DESCRIPTION

[0096] In order to make the purpose, technical scheme and advantages of the present application clearer, the technical scheme of the present application will be described clearly and completely below in combination with the drawings in the embodiments of the present application. The embodiments described in the present application are only a part of the embodiments of the present application, but not all the embodiments. Based on the spirit of the present application, other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the present application.

[0097] Traditional convolutional neural network learning model

[0098] The traditional convolutional neural network (CNN) has been applied to multiple industrial data analysis fields due to its powerful feature extraction and classification ability. The overall network structure of the CNN can be divided into two parts, one part is a feature extraction network composed of multiple convolutional layers, and the other part is a classifier based on the feature extraction result to divide the sample class attribution. Taking the commonly used softmax classification function as an example, the type attribution probability calculation of each sample feature is shown in formula (1):

[0099]

[0100] where w i and b i are the weight and bias parameters of the i i th class, C is the total number of sample types, f(·) represents the convolutional network function, y represents the class to which the sample x belongs, and the sum of the classification probabilities of all types in the sample set is 1, as shown in formula (2):

[0101]

[0102] As can be seen from the formula, the overall probability space is divided according to the total number of types in the sample set. If the sample can be divided into 4 types, then the probability space of each type is one quarter of the overall space, as shown in the left half part. Figure 2

[0103] ​The above classification problem analysis is called closed set classification, that is, the whole probability space is divided by known classes. If a sample that does not appear in the training set appears in the application process (i.e. an open set problem), since the classifier does not leave a corresponding probability space for unknown classes, the unknown sample will be confused in the probability space of known types. Therefore, the traditional closed set classification model does not have the ability to distinguish and distinguish unknown samples.

[0104] From the above analysis, the probability space division idea of the closed set classification model is the key to the problem of not being able to identify unknown samples. Therefore, this paper introduces a convolutional prototype network, discards the traditional equal probability interval idea, and divides the probability space by determining the class center point and neighborhood to process the classification of known samples. At the same time, the overall probability space is regarded as an open domain, and the relationship between unknown samples and known samples is defined by the distance between classes, so as to realize the distinction of unknown samples.

[0105] Embodiment 1.

[0106] A circuit breaker fault diagnosis method based on a convolutional prototype network. As shown in Figure 1 The steps include:

[0107] Step 1, collect circuit breaker fault data, and pre-process the circuit breaker fault data to obtain a circuit breaker fault data training sample set;

[0108] The data preprocessing method includes: denoising, filtering out noise in the data collection process, and the preferred denoising method in this embodiment includes: mean, EMD (Empirical Mode Decomposition), wavelet decomposition; regularizing data to ensure consistent data length.

[0109] As shown in Table 1, the circuit breaker fault data is divided into: coil aging data A, coil aging data B, coil aging data C, iron core sticking data A, and iron core sticking data B according to different fault types and different fault degrees.

[0110] Table 1 Fault data type

[0111]

[0112] In this preferred embodiment, without damaging the internal structure of the circuit breaker, two typical fault types are simulated: ① aging of the opening and closing coils; ② core jamming. The experimental object is the commonly used LW□-126 / T3150-40 circuit breaker (110kV). The equipment used in the experiment includes: digital oscilloscope, current clamp, sliding rheostat, and several counterweights. The degree of coil aging faults can generally be divided into slight aging, moderate aging, and severe aging. Different degrees of coil aging are simulated by connecting an adjustable resistor in series in the coil circuit, with the resistance value adjusted to 20%, 50%, and 80% of the actual coil resistance. In addition, core jamming faults are simulated by suspending different counterweights at the end of the core, with the degree of fault adjusted by the weight of the suspended weights, which are 0.03kg and 0.06kg respectively. The category labels for each specific fault type are shown in Table 1, and the current waveforms obtained from the simulation experiment are shown in Table 1. Figures 3-4 As shown.

[0113] In data processing, this embodiment preferably uses a waveform sampling rate of 16kHz, with each fault sample containing 800 current data points. Therefore, the sample format input to the prototype convolutional network is 1*800. Consistent with Table 1, the waveforms of the six types of fault samples are as follows: Figure 5 As shown.

[0114] Step 2: Construct a circuit breaker fault diagnosis model using a convolutional prototype network. Train the model using the input training sample set. Using prototype sample points of each type as cluster centers, use the spatial distance of sample features as the optimization objective of the circuit breaker fault diagnosis model using the convolutional prototype network. Use a classification function to classify the output features of the convolutional layer. Introduce class neighborhoods to train the loss function. The loss function is used to drive the update of the hyperparameters of the convolutional layer to effectively improve the intra-class clustering and inter-class dispersion of fault sample features, thereby improving the model's accuracy in classifying sample types.

[0115] As shown in Table 2, the circuit breaker fault diagnosis model of the convolutional prototype network includes: an input layer, a convolutional pooling layer 1, a convolutional pooling layer 2, and a fully connected layer.

[0116] Table 2 Prototype Convolutional Network Architecture

[0117]

[0118] like Figure 2 As shown in the right half, the circuit breaker fault diagnosis model of the convolutional prototype network selects several prototype samples {m} for each class during the training process. ij / j∈{1,2,…,K}} where i represents the number of categories i∈{1,2,…,C}, and K represents the number of prototype samples selected for each category, such as Figure 2The prototype sample is similar to the past memory of the human brain for the category, and the type judgment is realized by matching with the past memory.

[0119] Step 2.1, the classification function is constructed in the idea of clustering, if the sample has the highest matching degree relative to a certain class, that is, the sample distance is the smallest, the sample can be determined to belong to the category, and the classification function is shown as formulas (3)-(4):

[0120]

[0121]

[0122] In the formula,

[0123] C represents the total number of sample types,

[0124] i represents the number of categories,

[0125] j represents the prototype sample point number of each class. j∈{1,2,…,K} because the number of prototype sample points of each class in the present application is 1, that is, there is only one prototype sample point in a class.

[0126] K represents the number of prototype samples selected for each type,

[0127] f(x) represents a convolution network function,

[0128] m ij represents the prototype sample point.

[0129] g i (x) represents the matching score of sample x and known class i.

[0130] For the sample, if it does not match all known classes, it can be expressed as:

[0131]

[0132] In the formula,

[0133] T represents the matching degree threshold of sample features and prototype samples of each class.

[0134] Step 2.2, the size of the matching degree is quantified by a probability form, and the classification function is improved.

[0135] If the distance value is used as a method for quantifying the size of the sample feature matching degree, it may appear to be infinite, and the value range of formula (4) will reach (-∞, 0). Therefore, the size of the matching degree is quantified by a probability form, so as to facilitate the setting of the corresponding parameters in the subsequent analysis.

[0136] The matching degree calculation method is converted into formula (6) shown:

[0137]

[0138] In the formula,

[0139] represents the distance between two vectors,

[0140] p i (x) represents the matching score of the probability quantized sample x and the known class i.

[0141] γ represents a hyperparameter for controlling the conversion of distance results to probability form, and takes a value of 0.001,

[0142] Similarly, formula (5) is rewritten as:

[0143]

[0144] In the formula, T1 represents the matching degree threshold in the probability form.

[0145] Step 2.3, estimate the sample feature distance according to the improved classification function.

[0146] Assuming that for sample x, i1 and i2 represent two samples in a certain type whose feature distances are closest and second closest to sample x, and j is a non-same-class sample, then the feature distance relationship of the three samples is expressed as:

[0147] g i1 (x) > g i2 (x) > > g j (x) (8)

[0148] Then the sample feature distance approximation formula is shown as follows:

[0149]

[0150] Therefore, the selection of the matching degree threshold T1 can be obtained from formula (9).

[0151] Step 2.4, introduce class neighborhood to train the loss function, and directly drive the optimization of the convolutional prototype network for sample feature extraction with the goal of feature clustering.

[0152] In the definition of the training loss function, the training goal of the multi-classification model is to make the sample x belong to its correct class y, and not belong to all other classes c (c≠y). In order to convert the distance into binary classification, that is, "yes" or "no" belongs to this class, the model is set to be greater than 0 for each prototype sample point m ij Set a radius r to form a class neighborhood. The expression is as follows:

[0153]

[0154] wherein,

[0155] f represents the result of sample x after convolution network extraction,

[0156] m ij represents the prototype sample point,

[0157] r represents the radius,

[0158] R i represents the decision area.

[0159] For each class i, the class neighborhood constitutes the decision area R i of its feature space, and the center point of the decision area, i.e. the prototype sample point, is the mean point of the features of each sample in the class and is constantly updated as the features of the samples change in the training.

[0160] For sample x, f(x) ∈ R i indicates that x belongs to class i, otherwise it does not belong to class i, and if there is a sample that does not satisfy f(x) ∈ R i for all types, it can be judged as an unknown class. Through a one-to-many training strategy, the target is sample x, and its feature classification can be expressed as:

[0161]

[0162] wherein,

[0163] R y represents the decision area of sample class y,

[0164] R c represents the decision area of sample class c

[0165] The above calculation process of binary classification is expressed as a function form, and then:

[0166]

[0167]

[0168] wherein,

[0169] h i (x) represents the function expression form of the same class sample i,

[0170] h y (x) represents the result of sample class y,

[0171] h c (x) represents the result of sample class c (c ≠ y).

[0172] In order to make the loss function more smooth, the binary discriminant function is converted into a probability function (denoted as σ) by using the sigmoid function, that is:

[0173] σ i (x) = sigmoid(h i (x)) (14)

[0174] The loss function of the proposed model is:

[0175]

[0176] In the formula,

[0177] σ i (x) represents the probability result of the same sample i,

[0178] σ y (x) represents the probability result of sample class y,

[0179] σ c (x) represents the probability result of sample class c (c≠y),

[0180] In order to avoid setting the parameter r manually, this paper takes the radius r as a hyperparameter, that is, the maximum r value of the training data set is updated constantly in the training process. As can be seen from the loss function of formula (15), the goal of the overall model optimization is to make the intra-class distance of sample features closer and the inter-class distance farther, which directly drives the optimization of sample feature extraction by convolution network in essence with the goal of feature clustering.

[0181] Step 3: Input the circuit breaker fault data sample set into the circuit breaker fault diagnosis model of the trained convolution prototype network to obtain the fault diagnosis result.

[0182] Embodiment 2.

[0183] Based on the example of a circuit breaker fault diagnosis method based on a convolution prototype network according to embodiment 1.

[0184] In the example, the python3.7 based on the tensorflow deep learning framework is adopted, the hardware platform is configured as Intel i7-9700 processor, 64-bit WIN10 operating system, memory 32GB, and graphics card memory 6GB. The project team has carried out 100 experiments for each type in table 1, that is, F11~F31 a total of 6 kinds of situations, and the total amount of data is 600. In the verification process, the training data and test data are divided according to the ratio of 7:3.

[0185] Firstly, all types of circuit breaker fault data set are regarded as known classes, and the feature extraction ability of the proposed algorithm is analyzed. For example, Figure 6After 300 training, the model completed the learning of the characteristics of the training data. In the figure, the classification accuracy no longer changes after completing 30 iterations, while the loss value result is still slightly decreasing until it remains stable after completing 157 iterations, because the m i and r parameters are in the stable state, the class neighborhood of each fault type has been roughly determined, so the type attribution of the sample is basically determined, thereby making the classification accuracy remain unchanged. However, because the hyperparameters of the convolutional neural network are still in the fine-tuning state, the loss function continues to change slightly. Finally, when the hyperparameters of the model are stable and unchanged, the learning of the training data is completed.

[0186] Figure 7 The results of the feature extraction of the training data by the model proposed in this paper are as follows: Figure 8 The results of the feature extraction by the traditional CNN model (the traditional CNN network structure is consistent with Table 2, the output of the fully connected layer is 6-dimensional, and is compressed to 2-dimensional after principal component analysis). Figure 8 The white sample points in each class cluster are the prototype sample points m i It can be seen that each type of sample has m i as the class center (the center circle point), and the spatial distance between each sample is close, compared with Figure 8 The traditional CNN feature extraction effect, the class aggregation and class separation have been obviously improved, which is more conducive to the subsequent classifier for judging the sample type.

[0187] To more accurately reflect the feature extraction effect of the two models, the silhouette coefficient index is used to quantitatively analyze the results of the two models to evaluate the class aggregation and class dispersion of the features of each type of sample extracted by the two models. The calculation formula of the silhouette coefficient is as follows: Figure 9 9

[0188]

[0189] Among them, the average distance of sample i to other samples in the same cluster is a(i), which is called the intra-cluster dissimilarity. The average distance of sample i to C j cluster is b ij , where j is the cluster number. When the s value is closer to 1, it proves that the aggregation degree of the class sample is better, otherwise, it is worse. As shown in Figure 10 , the traditional CNN model has a certain degree of confusion in the extraction of the features of F22 and F23 two types of faults, the intra-class distance of some samples of F22 and F23 is greater than the inter-class distance, and the corresponding silhouette coefficient appears negative value (the minimum is-0.67). In contrast, the silhouette coefficient result of the proposed model is better, and the aggregation of the sample features is better. ​​

[0190] Furthermore, to verify the proposed model's ability to identify unknown fault types, this paper uses the first five types of fault data in the fault dataset as known classes (i.e., the training dataset) and the sixth type as unknown classes (which only appear in the test set). The effectiveness of the model is verified using the same ratio of training to testing data as in the aforementioned example. Figure 11 Feature extraction results for 5 types of training data. Figure 11 The feature extraction results for the six types of test data are shown, where the yellow highlighted feature points inside the red ellipse represent the fault type F32 that did not appear in the training data.

[0191] Depend on Figure 9 The inter-class distance of the unknown class is larger than that of the known class, indicating that the model exhibits ideal repulsion for unknown samples in the test set, effectively distinguishing them from the known class. However, regarding intra-class clustering, the features of the known class, after effective model learning, show better clustering, while the feature domains of the unknown class samples have a wider range and are more discrete. Nevertheless, this does not affect the model's effective identification of the unknown class. Subsequently, this paper uses the same training and testing method with five other types of fault data as the unknown class, and the results are similar to those described above, which will not be elaborated upon here.

[0192] To better demonstrate the superiority of the proposed algorithm, this paper compares it with the methods proposed in Method 1 and Method 2 using the same dataset. Unlike the CNN feature extraction network used in this paper, Method 1 employs a Deep Belief Network (DBN) for feature extraction. Then, based on transfer learning, it assists the model in judging measured samples by learning from simulation data, effectively improving the model's learning ability in small sample situations. However, it still uses a softmax classification model in the classification stage; therefore, this algorithm can only judge known classes. Method 2, on the other hand, extracts eight feature phasors {I1,I2,I3,t1,t2,t3,t4,t5} from the opening and closing current waveforms based on signal processing algorithms in the feature extraction stage. Then, it uses the similarity matrix of sample features as the basis for clustering, and finally, through cluster analysis, it achieves the diagnosis of known and unknown class faults. The comparison results are shown in Table 3.

[0193] This paper compares and analyzes the application effects of three methods on the dataset using three dimensions: the average and minimum values ​​of the silhouette coefficients, and the recognition accuracy. The average silhouette coefficient examines the overall clustering of features for this type of sample, while the minimum value directly reflects whether the feature domain of this type of sample overlaps with other types in space. ​ The overlap between F22 and F23 is a concern. If this overlap occurs, the classifier's accuracy in recognizing that type of sample will be affected.

[0194] Table 3 Comparison of analysis results of three methods

[0195]

[0196] In the analysis of the known class (closed set problem) for the six classes, both method one and method two have confusion problems for the analysis of F23 type. It can be seen that when the difference of sample characteristics is not obvious, the expression ability of the two methods for the sample is also weak. Among them, method one has the confusion of F22 and F23 classes, and the feature aggregation of other types is good. However, due to the use of the traditional 8 feature method, the feature dispersion is relatively large, so that the average value of the contour coefficient of other types is relatively low in addition to the confusion of F22 and F23 classes. Compared with the first two methods, the method proposed in this paper achieves good feature separation for each type of sample, which is because the prototype sample point as the clustering center well aggregates the feature distribution of the in-class samples.

[0197] In addition, only method two and method three can achieve unknown type sample recognition. The affinity propagation clustering algorithm used by method two has a relatively strict requirement for the spatial distribution of samples. Due to the relatively discrete distribution of the feature similarity result of individual sample points, and the small difference between different types of samples, the method has 3 misclassified samples in the unknown class recognition process (the total number of test samples is 30). Moreover, the method is prone to error loss in the feature extraction process, and the result will also affect the similarity calculation of the sample. In contrast, the method proposed in this paper is more effective in feature extraction, avoids the cumbersome signal processing feature extraction step, realizes the recognition of unknown classes, and is easier to apply to engineering practice.

[0198] Embodiment 3.

[0199] A circuit breaker fault diagnosis system based on a convolution prototype network, comprising: a collection module, a modeling and training module, and a diagnosis module, wherein:

[0200] The collection module is used to collect circuit breaker fault data, and to pre-process the circuit breaker fault data to obtain a circuit breaker fault data training sample set;

[0201] The modeling and training module is used to construct a circuit breaker fault diagnosis model of the convolution prototype network, to input the training sample set to train the model, and to take the sample feature space distance as an optimization target of the circuit breaker fault diagnosis model of the convolution prototype network;

[0202] The diagnosis module is used to input the circuit breaker fault data sample set into the trained circuit breaker fault diagnosis model of the convolution prototype network to obtain a fault diagnosis result.

[0203] Embodiment 4.

[0204] Embodiment 4 of the present application provides a computer readable storage medium.

[0205] A computer readable storage medium, having a program stored thereon, which program, when executed by a processor, implements the steps in the circuit breaker fault diagnosis method based on a convolution prototype network according to Embodiment 1 of the present application.

[0206] The detailed steps are the same as those in the circuit breaker fault diagnosis method based on a convolution prototype network according to Embodiment 1, which will not be repeated here.

[0207] Embodiment 5.

[0208] Embodiment 5 of the present application provides an electronic device.

[0209] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor, and the processor implements the steps in the circuit breaker fault diagnosis method based on a convolution prototype network according to Embodiment 1 of the present application when executing the program.

[0210] The detailed steps are the same as those in the circuit breaker fault diagnosis method based on a convolution prototype network according to Embodiment 1, which will not be repeated here.

[0211] The present application has the beneficial effect that, compared with the prior art, the present application proposes a circuit breaker fault diagnosis method based on a convolution prototype network. First, the fault data is preprocessed, and the normal data and the fault sample data are combined to highlight the difference of the fault, and then the traditional closed set learning model is improved, and the classification function is constructed with the idea of clustering to adapt to the processing of the open data set, and the diagnosis of unknown circuit breaker fault types is realized. At the same time, taking each type of prototype sample point as the clustering center, and taking the sample feature space distance as the optimization target of the convolution feature self-extraction network, the intra-class aggregation and inter-class dispersion of the fault sample features are effectively improved, and the classification accuracy of the model to the sample type is improved. In this paper, the simulation experimental data is taken as the analysis object, the accuracy of the proposed method in feature extraction and unknown type diagnosis is analyzed in detail, and the proposed method is compared with the existing method, and the superiority of the model in the feature extraction and classification stages is analyzed.

[0212] The present disclosure can be a system, a method, and / or a computer program product. The computer program product can include a computer readable storage medium having computer readable program instructions embodied therewith, wherein the computer readable program instructions are used to cause a processor to implement various aspects of the present disclosure.

[0213] Computer readable storage media can be tangible storage media which can retain and store instructions for use by an instruction execution device. Computer readable storage media can be, for example, but is not limited to, an electronic storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer readable storage media include the following: a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), a static random access memory (SRAM), a portable compact disc read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanically encoded device such as punch-cards or raised structures in a groove having instructions recorded thereon, and any suitable combination of the foregoing. A computer readable storage medium, as used herein, is not to be construed as being transitory signals per se, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through a waveguide or other transmission media (e.g., light pulses passing through a fiber-optic cable), or electrical signals transmitted through a wire.

[0214] Computer readable program instructions described herein can be downloaded to respective computing / processing devices from a computer readable storage medium or to an external computer or external storage device via a network, for example, the Internet, a local area network, a wide area network and / or a wireless network. The network can comprise copper transmission cables, optical transmission fibers, wireless transmission, routers, firewalls, switches, gateway computers and / or edge servers. A network adapter card or network interface in each computing / processing device receives computer readable program instructions from the network and forwards the computer readable program instructions for storage in a computer readable storage medium within the respective computing / processing device.

[0215] Computer readable program instructions for carrying out operations of the present disclosure can be assembler instructions, instruction-set-architecture (ISA) instructions, machine instructions, machine dependent instructions, microcode, firmware instructions, state-setting data, or either source code or object code written in any combination of one or more programming languages, including an object oriented programming language such as Smalltalk, C++ or the like and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The computer readable program instructions can execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server. In the latter scenario, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection can be made to an external computer (for example, through the Internet using an Internet Service Provider). In some embodiments, electronic circuitry including, for example, programmable logic circuitry, field-programmable gate array (FPGA), or programmable logic array (PLA) can execute the computer readable program instructions by utilizing state information of the computer readable program instructions to personalize the electronic circuitry, in order to perform aspects of the present disclosure.

[0216] The computer readable program instructions can also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable apparatus or other device to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide processes for implementing the functions / acts specified in the flowchart and / or block diagram block or blocks.

[0217] The computer readable program instructions can also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable apparatus or other device to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide processes for implementing the functions / acts specified in the flowchart and / or block diagram block or blocks.

[0218] The computer readable program instructions can also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus or other device to produce a computer implemented process such that the instructions which execute on the computer, other programmable data processing apparatus, or other device implement the functions / acts specified in the flowchart and / or block diagram block or blocks.

[0219] The computer readable program instructions can also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus or other device to produce a computer implemented process such that the instructions which execute on the computer, other programmable data processing apparatus, or other device implement the functions / acts specified in the flowchart and / or block diagram block or blocks.

[0220] Finally, it should be noted that the above-described embodiments are merely intended for describing and illustrating, not limiting, the technical solutions of the present application. Although the present application has been described in detail with reference to the above embodiments, it should be understood by those skilled in the art that the specific embodiments of the present application can be modified or replaced equivalently without departing from the spirit and scope of the present application, and any modification or equivalent replacement without departing from the spirit and scope of the present application should be covered within the protection scope of the claims of the present application.

Claims

1. A circuit breaker fault diagnosis method based on a convolutional prototype network, characterized by, The method comprises the following steps: Step 1, collecting circuit breaker fault data, and pre-processing the circuit breaker fault data to obtain a circuit breaker fault data training sample set; Step 2, constructing a circuit breaker fault diagnosis model of a convolutional prototype network, inputting the training sample set to train the model, and taking the sample feature space distance as the optimization target of the circuit breaker fault diagnosis model of the convolutional prototype network; The circuit breaker fault diagnosis model of the convolutional prototype network in step 2 comprises an input layer, a convolutional pooling layer 1, a convolutional pooling layer 2 and a fully connected layer, a classification function is used to classify the convolutional layer output features, a class neighborhood is introduced to train the loss function, the loss function is used to drive the update of the convolutional layer hyperparameters, and the model training process specifically comprises: Step 2.1, constructing a classification function in a clustering method, classifying the circuit breaker fault data training sample set, and determining the sample category if the matching degree of the sample relative to a certain category is the highest, that is, the sample feature space distance is the smallest; Step 2.2, quantifying the size of the matching degree in a probabilistic form to improve the classification function; Step 2.3, estimating the sample feature space distance according to the improved classification function; Step 2.4, introducing a class neighborhood to train the loss function to directly drive the optimization of the circuit breaker fault diagnosis model of the convolutional prototype network to sample feature extraction; Step 2.4 specifically comprises: in the definition of the training loss function, the training target of the multi-classification model is to make the sample belong to its correct category, and not belong to all other categories; in order to convert the distance into a binary classification, that is, whether it belongs to this category, a radius is set for each prototype sample point of the model to form a class neighborhood, and the expression is as follows: In the formula, f represents a result after a sample x is extracted by a convolution network, m ij represents a prototype sample point, r represents a radius, R i represents a decision region; For each class i, the class neighborhood constitutes the decision region R in its feature space i The center point of the decision region, i.e. the prototype sample point, is the mean point of the features of the samples in the class and is updated constantly as the features of the samples change during training. For sample x, f(x) ∈ R i x belongs to i class, otherwise not belong to i class, if there is a sample i x, then its feature classification can be expressed as: In the formula, R y represents the decision region of the sample class y, R c represents the decision region of the sample class c, and the calculation process of the above binary classification is expressed in the form of a function as follows: where h i (x) represents the functional form of the homogenous sample i, h y (x) represents the result of the sample class y, h c (x) represents the result of the sample class c, where c≠y; In order to make the loss function smoother, the sigmoid function is used to convert the binary discriminant function into a probability function, denoted as σ, as follows: σ i (x) = sigmoid(h i (x)) (5) The loss function of the model is: where σ i (x) represents the probability outcome for a sample of the same class i, σ y (x) represents the probability outcome for a sample of class y, σ c (x) represents the probability outcome for a sample of class c, where c≠y; Step 3, inputting the circuit breaker fault data sample set into the trained circuit breaker fault diagnosis model of the convolutional prototype network to obtain a fault diagnosis result.

2. The circuit breaker fault diagnosis method based on the convolutional prototype network according to claim 1, wherein the circuit breaker fault data in step 1 comprises coil aging data A, coil aging data B, coil aging data C, iron core jamming data A and iron core jamming data B.

3. The circuit breaker fault diagnosis method based on the convolutional prototype network according to claim 1, wherein the circuit breaker fault data preprocessing method in step 1 comprises a mean value method, an EMD method and a wavelet decomposition method.

4. The circuit breaker fault diagnosis method based on the convolutional prototype network according to claim 1, wherein step 2.1 comprises that the classification function is as shown in the formula: In the formula, C represents the total number of sample types, i represents the sample classification category, j represents the prototype sample point number of each category, j∈{1,2,…,K}, K represents the number of prototype samples selected for each type, f(x) represents a convolutional prototype network function, for a sample, if it does not match all known categories, it is expressed as: In the formula, T represents the matching degree threshold of the sample features and the prototype samples of each category. m ij denotes a prototype sample point, g i (x) represents the match score of sample x to known class i; ​ ​ ​ 5. The circuit breaker fault diagnosis method based on the convolutional prototype network according to claim 4, characterized in that, Step 2.2 includes: the matching degree calculation method is converted into the formula shown in the formula: In the formula, denotes the distance between two vectors, p i (x) denotes the match score of the probability-quantized sample x with the known class i, γ represents a hyperparameter for converting the control distance result into a probability form, and takes a value of 0.001, The formula (9) is rewritten as: In the formula, T1 represents a matching degree threshold value in the probability form.

6. The circuit breaker fault diagnosis method based on the convolutional prototype network according to claim 5, characterized in that, Step 2.3 specifically includes: assuming that, for a sample x, i1 and i2 represent two samples closest to and second closest to the sample x in a certain type of feature distance, and j is a non-same sample, then the feature distance relationship of the three samples is expressed as: g i1 (x) > g i2 (x) g j (x) (12) Then the sample feature distance approximation estimation formula is shown as follows: Therefore, the selection of the matching degree threshold value T1 is obtained from the formula (13).

7. A circuit breaker fault diagnosis system based on convolutional prototype network using the method of any one of claims 1-6, comprising: The acquisition module, the modeling and training module, and the diagnosis module are characterized in that: The acquisition module is used to collect circuit breaker fault data, and pre-process the circuit breaker fault data to obtain a circuit breaker fault data training sample set; The modeling and training module is used to construct a circuit breaker fault diagnosis model of the convolutional prototype network, input the training sample set to train the model, and take the sample feature space distance as an optimization target of the circuit breaker fault diagnosis model of the convolutional prototype network; The diagnosis module is used to input the circuit breaker fault data sample set into the trained circuit breaker fault diagnosis model of the convolutional prototype network to obtain a fault diagnosis result.

8. A terminal comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is used to operate according to the instructions to perform the steps of the circuit breaker fault diagnosis method based on the convolutional prototype network according to any one of claims 1-6.

9. A computer readable storage medium having stored thereon a computer program, characterized in that The program is executed by the processor to implement the steps of the circuit breaker fault diagnosis method based on the convolutional prototype network according to any one of claims 1-6.

Images